CVE-2025-14926 affects the Red Hat AI Inference Server Model Optimization Tools 3.3.5 (CUDA) through a code injection vulnerability in the Hugging Face Transformers library's convert_config function. This function fails to validate a user-supplied string before executing it as Python code, enabling arbitrary code execution when processing malicious SEW model checkpoints. Exploitation requires user interaction to process a specially crafted checkpoint file. Due to the typical privilege level of users processing these files and the need for user interaction, the severity is rated as important by Red Hat. No patch or official fix is currently available according to the vendor advisory. Mitigation involves avoiding untrusted model checkpoints or using sandboxing techniques.

Successful exploitation can lead to arbitrary code execution in the context of the user running the conversion process, potentially compromising data integrity, confidentiality, and availability within that user context. However, the need for user interaction and the typical lack of administrative privileges reduce the risk of full system compromise. The vulnerability is rated with high impact on confidentiality, integrity, and availability within the user scope.

No official fix is currently available for this vulnerability. Users should avoid converting SEW model checkpoints from untrusted or unverified sources. It is recommended to ensure all processed model checkpoints originate from trusted repositories or have verified integrity. If processing untrusted models is necessary, it should be done within isolated sandbox environments to limit potential damage.